Re: [Patch v5 1/4] memory: tegra: Add memory controller channels support

[Dmitry Osipenko <dmitry.osipenko@collabora.com>]

On 3/25/22 07:50, Ashish Mhetre wrote:
> 
> 
> On 3/19/2022 9:12 PM, Dmitry Osipenko wrote:
>> External email: Use caution opening links or attachments
>>
>>
>> 16.03.2022 12:25, Ashish Mhetre пишет:
>>>  From tegra186 onwards, memory controller support multiple channels.
>>> Add support for mapping address spaces of these channels.
>>> Make sure that number of channels are as expected on each SOC.
>>> During error interrupts from memory controller, appropriate registers
>>> from these channels need to be accessed for logging error info.
>>>
>>> Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
>>> ---
>>>   drivers/memory/tegra/mc.c       |  6 ++++
>>>   drivers/memory/tegra/tegra186.c | 52 +++++++++++++++++++++++++++++++++
>>>   drivers/memory/tegra/tegra194.c |  1 +
>>>   drivers/memory/tegra/tegra234.c |  1 +
>>>   include/soc/tegra/mc.h          |  7 +++++
>>>   5 files changed, 67 insertions(+)
>>>
>>> diff --git a/drivers/memory/tegra/mc.c b/drivers/memory/tegra/mc.c
>>> index bf3abb6d8354..3cda1d9ad32a 100644
>>> --- a/drivers/memory/tegra/mc.c
>>> +++ b/drivers/memory/tegra/mc.c
>>> @@ -749,6 +749,12 @@ static int tegra_mc_probe(struct platform_device
>>> *pdev)
>>>        if (IS_ERR(mc->regs))
>>>                return PTR_ERR(mc->regs);
>>>
>>> +     if (mc->soc->ops && mc->soc->ops->map_regs) {
>>> +             err = mc->soc->ops->map_regs(mc, pdev);
>>> +             if (err < 0)
>>> +                     return err;
>>> +     }
>>> +
>>>        mc->debugfs.root = debugfs_create_dir("mc", NULL);
>>>
>>>        if (mc->soc->ops && mc->soc->ops->probe) {
>>> diff --git a/drivers/memory/tegra/tegra186.c
>>> b/drivers/memory/tegra/tegra186.c
>>> index 3d153881abc1..a8a45e6ff1f1 100644
>>> --- a/drivers/memory/tegra/tegra186.c
>>> +++ b/drivers/memory/tegra/tegra186.c
>>> @@ -139,11 +139,62 @@ static int tegra186_mc_probe_device(struct
>>> tegra_mc *mc, struct device *dev)
>>>        return 0;
>>>   }
>>>
>>> +static int tegra186_mc_map_regs(struct tegra_mc *mc,
>>> +                             struct platform_device *pdev)
>>> +{
>>> +     struct device_node *np = pdev->dev.parent->of_node;
>>> +     int num_dt_channels, reg_cells = 0;
>>> +     struct resource *res;
>>> +     int i, ret;
>>> +     u32 val;
>>> +
>>> +     ret = of_property_read_u32(np, "#address-cells", &val);
>>> +     if (ret) {
>>> +             dev_err(&pdev->dev, "missing #address-cells property\n");
>>> +             return ret;
>>> +     }
>>> +
>>> +     reg_cells = val;
>>> +
>>> +     ret = of_property_read_u32(np, "#size-cells", &val);
>>> +     if (ret) {
>>> +             dev_err(&pdev->dev, "missing #size-cells property\n");
>>> +             return ret;
>>> +     }
>>> +
>>> +     reg_cells += val;
>>> +
>>> +     num_dt_channels =
>>> of_property_count_elems_of_size(pdev->dev.of_node, "reg",
>>> +                                                       reg_cells *
>>> sizeof(u32));
>>> +     /*
>>> +      * On tegra186 onwards, memory controller support multiple
>>> channels.
>>> +      * Apart from regular memory controller channels, there is one
>>> broadcast
>>> +      * channel and one for stream-id registers.
>>> +      */
>>> +     if (num_dt_channels < mc->soc->num_channels + 2) {
>>> +             dev_warn(&pdev->dev, "MC channels are missing, please
>>> update\n");
>>
>> Update what?
>>
>>> +             return 0;
>>> +     }
>>> +
>>> +     mc->mcb_regs = devm_platform_get_and_ioremap_resource(pdev, 1,
>>> &res);
>>
>> Can't we name each reg bank individually in the DT and then use
>> devm_platform_ioremap_resource_byname()?
>>
> That can be done but I think current logic will be better as we can
> simply ioremap them by running in loop and assigning the mc_regs array.
> Otherwise there will be like 17 ioremap_byname() individual calls for
> Tegra194 and Tegra234.
> Will it be fine having that many ioremap_byname() calls?
> Also, Tegra186 has 5 channels which are less than Tegra194 and Tegra234.
> If we go with ioremap_byname() then we'll have to differentiate number
> of ioremap_byname() calls.
for (i = 0; i < mc->soc->num_channels; i++) {
	sprintf(name, "mc%u", i);
	err = devm_platform_ioremap_resource_byname(dev, name);
	...
}